Garmin Connect MCP

Remote MCP server for Garmin Connect. Gives AI assistants access to your Garmin health and activity data over the internet — no local server or app required.

Built on top of Taxuspt/garmin_mcp and cyberjunky/python-garminconnect.

What makes this different from Taxuspt/garmin_mcp

Taxuspt’s original server runs locally on your machine and requires a locally installed AI desktop client. This version runs as a cloud service, protected by GitHub OAuth, so it works from any MCP-compatible AI assistant on any device — without running anything locally.

What it does

Exposes ~150 Garmin Connect tools so AI assistants can query your health data directly:

• Health & Wellness: sleep, HRV, stress, body battery, heart rate, SpO2, respiration, recovery time
• Activities: runs, rides, swims — with splits, weather, HR zones, FIT file analysis
• Training: readiness, status, VO2max, load, running tolerance, acclimation, trends
• Body composition, weight, hydration, energy balance
• Workouts (incl. builders and Garmin Coach), gear, nutrition, courses, calendar events, challenges

The tool modules are kept in sync with Taxuspt/garmin_mcp; this repo adds the remote server layer (GitHub OAuth, persistent token stores, filesystem-tool block).

Choosing which tools are exposed

Every tool definition is sent to the AI assistant with each conversation, so ~150 tools cost noticeably more context than a focused set. Use GARMIN_ENABLED_TOOLS (allowlist) or GARMIN_DISABLED_TOOLS (denylist) — comma-separated tool names — to expose only what you need. Unknown names are logged at startup.

Blocked filesystem tools

download_activity_file, download_course_gpx, set_fit_download_dir and upload_course read or write arbitrary paths on the server. That is fine for a local single-user install, but on a remote server one (possibly prompt-injected) call could overwrite the OAuth token store, the Garmin token or the application code. They are never registered here, whatever the tool filter says.

How it works

AI assistant → /authorize → GitHub login (+ 2FA) → /auth/callback
             → username verified → MCP access token issued → MCP connection

Access is protected by GitHub OAuth — only the GitHub account set in GITHUB_ALLOWED_USER_ID can authenticate. GitHub login with 2FA is required once every 30 days; tokens are persisted to a token store (file or Redis). No credentials are stored in the AI assistant’s configuration.

1. The AI assistant detects the MCP server requires OAuth
2. A browser window opens — you log in to GitHub with 2FA
3. The server verifies your GitHub account matches GITHUB_ALLOWED_USER_ID
4. The AI assistant receives a 30-day access token and a 30-day refresh token

Restart note: OAuth tokens are persisted to the token store, so